{ "source": "blender_high_quality compact task/code pairs, repacked v2", "source_dataset_audit": "paired_sft_data_20260926_v2/audit.json", "seed": 20260929, "packing": "one 128-byte model context; loss only on code suffix after newline Code newline", "selection": "unique text windows; no replacement; <=5 train rows/family and <=2/task pair", "min_supervised_code_bytes": 32, "candidate_counts": { "train_and_new_test": 40886, "validation": 6685, "heldout_test_families": 131 }, "split_sizes": { "train": 3200, "validation": 256, "test": 256 }, "family_counts": { "train": 1087, "validation": 80, "test": 82 }, "family_overlap": { "train_validation": [], "train_test": [], "validation_test": [] }, "code_target_length_bins": { "train": { "64-96": 1139, "32-63": 982, "97+": 1079 }, "validation": { "97+": 88, "32-63": 80, "64-96": 88 }, "test": { "64-96": 77, "32-63": 83, "97+": 96 } }, "sha256": { "train.jsonl": "822ac39227bba96fbab71f0a1c06a1de7eb5610b62f57aa4177e929a24dc3d00", "validation.jsonl": "e453818d0f6060950f3c32f96a9841f1c8d6879425dc1eeca545bdc040fa3b0e", "test.jsonl": "6e57ba289cf8f1171fe763bc5c83db454c482406427145656102fcd4a0d04b92" }, "limitations": [ "Source quality checks establish JSON/AST syntax and bpy imports, not Blender runtime or visual correctness.", "Unique means unique sampled windows in this run; historical overlap with every prior model training corpus cannot be proven.", "The held-out test families are carved deterministically from the source's original train families; test is not used for checkpoint selection." ] }